feat: add endpoint-isolated common warm-up screening

[tappe9]

Summary

Refs #73. Adds the separately versioned, non-publishing common-warmup-screen-v1 protocol; the initial JVM diagnostic and its interpretation remain untouched.

• Java / Node / Go × three endpoints × two independent clean starts: 18 fresh API/PostgreSQL traces, with correctness checked on three separate sacrificial instances
• Five seconds of initial load plus eight 30-second windows; fixed cumulative-load candidate grid 5/35/65 seconds against disjoint reference windows 6–8
• Predeclared 5% throughput / 10% p95 bounds across candidate and later windows; every implementation, endpoint and both starts required; no candidate means inconclusive
• Preserves raw state, database fixture/identity, readiness, CPU/memory, timing/gaps, failure and cleanup evidence
• Separate manual-only trusted-main/read-only workflow: 120-minute job cap, 104-minute workload deadline, always-upload evidence; no automatic retries/publication

Configured load is 73.5 minutes; approximately 90–110 minutes overall is an estimate. Candidates would still require independent confirmation; nine-implementation runs, official profiles/cohorts and publication are outside this change.

Verification

• TDD for scheduler, failure paths, identity guards, environment, workflow and inclusive floating-point boundaries
• Independent whole-change review: no remaining findings after verified numeric-boundary fix
• Optimized Python benchmark suite: 257 tests passed
• Shared HTTP contract unit suite: 38 tests passed
• Workflow suite: 24 tests passed; actionlint passed
• Pinned Ruff 0.16.5 lint and formatting checks passed; README generator check and diff whitespace check passed
• Local full make test is blocked by missing Docker; the site browser test is blocked by this container's Chromium socket restrictions. These runtime/Compose/browser gates must pass in hosted exact-head CI before merge

The one additional diagnostic is separately user-authorized but has not been dispatched. Merge and successful post-merge checks precede that one run and subsequent artifact audit/Issue #73 report.